Cross-domain transfer learning algorithm for few-shot ship recognition in remote-sensing images
Cross-domain transfer learning aims to utilize public datasets as source data to improve the recognition accuracy of target data,breaking through the limitation that the category space between source data and target data must be consistent.For the few-shot remote-sensing ship recognition task,existing cross-domain transfer learning algorithms have the disadvantages of transfer category restriction and negative transfer effect.Therefore,a cross-domain transfer learning algorithm based on source data correlation sorting was proposed to solve the above problems.First,the target data were added reversely into the source domain recognition task.According to the variation of the source data recognition accuracy before and after the target data were added,various source data were classified into strong/weak/negative correlation samples,and only the strong correlation samples would be selected.Then,the self-supervised joint learning strategy was adopted to introduce the auxiliary self-supervised angle prediction branch into the classification network in the target domain.The selected strong correlation source samples were added but only into the training of the self-supervised branch,which avoided changing the main classification network structure.Randomly selecting 65 category samples as the source data from miniImageNet and conducting comparative experiments on few-shot ship targets in remote-sensing images yields the following results:1)When Resnetl8 is chosen as the classification network,the performance of the proposed algorithm is better than that of the Fine-tune algorithm,which is widely used in cross-domain transfer learning.Moreover,compared with the recognition algorithm,which only usesthe main classification network,the proposed algorithm improves the recognition accuracy of target data from 78.89％to 96.48％.2)Using different networks to sort correlations for the source data,the selected strong correlation source samples are not exactly the same and their degree of category coincidence is close to 60％.However,they are all helpful to the classification task of the target domain.At the end of this paper,through visualizing the extracted target features,it is verified that the target features extracted by using the proposed algorithm are more abundant and have higher generalization ability.The proposed algorithm has two main advantages.First,the weak/negative correlation source samples are eliminated by correlation sorting,which can avoid the occurrence of negative transfer effect.Second,by introducing the self-supervised angle prediction branch,the information of the strong correlation source samples is fully utilized and the features with more generalization ability are extracted while maintaining the structural integrity of the main classification network.
remote sensingship recognitionFew-Shot learningcross-domain transfer learningcorrelation sortingself-supervised learning
NETL
NSTL
万方数据
